Factors related to breast cancer detection mode and time to diagnosis in Alberta, Canada: a population-based retrospective cohort study

A retrospective breast cancer cohort of female residents of Alberta, who were diagnosed with a histologically confirmed first-ever primary breast cancer (International Classification of Disease for Oncology (ICD-O) 3rd edition code C50 behaviors 2 and 3 [12]) in years 2004 to 2010, was identified from the Alberta Cancer Registry, a population based cancer registry recognized for data completeness by the North American Association of Central Cancer Registries. All patients were included unless data to calculate outcome measures, as defined below, were missing.

During the study period, the province of Alberta had a population of about 3 to 3.5 million spread over 662,000 km² [13]. Approximately 56 % of the total population live within the two urban regions, 27 % live in small and medium regional cities, and roughly 17 % live in rural and remote areas [14]. Alberta has a publicly-funded provincially operated single-payer health care system, in which all residents have free access to standard medical care. The majority of physicians are remunerated via fee-for-service. Typically women are referred for mammography by their primary care physician (PCP) but screen-eligible patients can self-refer.

The following patient characteristics were obtained from the Alberta Cancer Registry: age, postal code of residence, regional health authority (RHA) of residence, disease stage, histologic grade and date of diagnosis. Disease stage was based on the American Joint Committee on Cancer staging [15]. Histology grade was based on the ICD-O morphology grade that represents the degree of differentiation of the tumor.

The postal codes of patients were used to obtain neighbourhood-level variables developed by Statistics Canada. Specifically, neighbourhood income level (in quintiles, QAIPPE) and an urban/rural variable (CSIZEMIZ) were obtained by linking patient postal codes to the 2006 Canadian Census Data [16]. The latter variable is defined by categorizing communities with populations less than 10,000 that have no influence from larger cities as rural and all other communities as urban.

International Classification of Disease, Tenth Revision, Clinical Modification [ICD-10-CM] diagnosis codes from three provincial administrative health care databases were used to calculate the Charlson Comorbidity Index: 1) the Ambulatory Care Classification System, all outpatient visits to hospitals in the province; 2) the Discharge Abstract Database, all inpatient hospital admissions in the province, and 3) the Physician Claims Database, all fee-for-service physician visits. All relevant codes in the period 30 months prior to breast cancer diagnosis were used to calculate the modified Charlson Comorbidity Index, [17, 18]. Codes for primary or metastatic cancer were excluded from calculations.

The Physician Claims Database was used to identify visits to a PCP using the ‘provider type’ code to calculate the Usual Provider Continuity (UPC) score. The UPC is calculated by dividing the number of visits to the PCP the patient has seen the most by the total number of all PCP visits a patient has had in a given time period [19, 20]. In order to capture typical PCP utilization, all PCP visits from 6 to 30 months prior to cancer diagnosis were included in the UPC calculation and a minimum of three PCP visits was required to calculate the UPC. Patients with less than three PCP visits in the period were classified as having “minimum PCP visits”.

The databases and datasets used in the study are not publicly available. The Alberta Cancer Registry data were made available upon ethics approval. The provincial administrative databases are governed by Alberta Health Services (AHS) via permission from the provincial ministry, Alberta Health (AH). AHS provided the provincial administrative data required for the study after reviewing the study protocol, receiving a signed confidentiality agreement (from MW) and receiving proof of ethics approval. Ethics approval for the study was obtained from the University of Alberta Health Research Ethics Board.

The outcome measures are percent screen-detected cancers and the length of the diagnostic interval. All breast cancer related diagnostic procedures, including screening/diagnostic mammograms, breast ultrasound and breast biopsies, were identified from the Physician Claims Database and the Screen Test Database. These two databases are complementary, capturing all breast cancer related screening and diagnostic procedures in the province from fee-for service and salaried radiologists, respectively. All breast-related procedures within a validated look-back period [21] from the date of diagnosis in the Alberta Cancer Registry were obtained.

Breast cancer was defined as screen-detected if the patient had a screening mammogram within the look-back period. The length of the diagnostic interval was defined as the time from the date of the screening mammogram to the date of breast cancer diagnosis, typically the first positive percutaneous or surgical biopsy date. The remaining cancers were defined as symptom-detected. The earliest breast cancer related diagnostic procedure in the look-back period, usually a diagnostic mammogram, was defined as the first relevant diagnostic test. The most proximal visit to a PCP within 6-months prior to the first relevant test was defined as the start of the diagnostic interval for symptom-detected patients, because in most cases a diagnostic test can only be conducted if a referral is made by a PCP. The diagnostic interval for symptom-detected patients, therefore, is defined as the time interval from the date of the PCP visit to the date of cancer diagnosis. Figure 1 summarizes the diagnostic interval by detection method as described above.

Data were linked using the unique provincial healthcare identification number which was anonymized for data analysis. Quality assurance and cross checks were performed on data sets during and after data linkage to ensure accuracy and completeness. Ethical approval was obtained from the Human Research Ethics Board at the University of Alberta before conducting the study.

Demographic, clinical and healthcare system factors were assessed for their univariate associations with each of the outcome measures. Chi-square tests were calculated to assess associations with percent screen-detected breast cancers and Kruskal-Wallis tests were calculated to assess associations with the length of the diagnostic interval. The demographic factors evaluated were age and neighbourhood income quintile. Patients were grouped as follows: 39 and under, 40–49, 50–69, 70–74, and 75 and over. The age groupings were based on screening guidelines and clinical practice in Alberta: 50–69 is the screen-eligible age group during the study period [22]; women aged 70–74 are included in the most recent breast screening guideline [23]; and women aged 40–49 years old can be referred for breast cancer screening in practice. Clinical factors included cancer stage at diagnosis, histology grade, and Charlson comorbidity index. Healthcare system factors included time period of cancer diagnosis, Usual Provider Continuity score (UPC), regional health authority (RHA), and urban/rural residence. The cut point for the time period variable was chosen based on the observed trend of percent screen-detected cancers over the years. Time period 1 includes cancers diagnosed in years 2004 to 2006 and time period 2 includes cancers diagnosed in years 2007 to 2010.

Multivariable logistic regression was used to examine the association of factors on the likelihood of cancer being screen-detected. The most parsimonious model was identified and presented as the final model. The interaction of time period by RHA was hypothesized and tested; the interaction was statistically significant and kept in the final model. For the diagnostic interval outcome, multivariable quantile regression models were fitted to estimate the effects of individual factors and the interaction effects between RHA and time period for screen and symptom detected patients separately. Briefly, quantile regression is similar to linear regression; it differs by modeling the median or specific quantiles defined by the user rather than the mean. For non-symmetric distributions such as time intervals, quantile regression is more appropriate than linear regression. The estimates from quantile regression are the difference in diagnostic interval (days) associated with each factor, making it simple to understand. Quantile regression models were run for the median and the 90^th percentile; the 90^th percentile was intended to represent diagnostic delay. All analyses were conducted using SAS 9.4 (SAS Institute, Cary NC).

There were 12,813 first-ever female breast cancers diagnosed in Alberta residents in years 2004 to 2010; 440 (3.4 %) were excluded from the study because of missing data that prevented the assignment of detection mode. Table 1 gives the descriptive statistics of the demographic, clinical, and healthcare system factors of the breast cancer patient cohort by detection mode. About 38 % of the cancer cases were screen-detected overall. Roughly 50 % of patients were in the screen-eligible age group (50 to 69 years of age), of whom 47 % were screen-detected. In the whole cohort, the majority of screen-detected patients, 72 %, had in situ or stage I cancer compared to only 39 % in the symptom-detected patients (P < 0.001). Similarly, screen-detected cancers had a lower histological grade than the symptom-detected cancers; 25 % of screen-detected tumors were low grade compared to 16 % of symptom detected (P < 0.001). About 57 % of screen-detected in situ cancers, however, were high grade (nuclear) compared to 50 % of those symptom-detected (P = 0.01, data not shown). Similar contrasts by detection mode were found on stage and histology grade among screen-eligible patients. Patients with a high UPC score were more likely to be screen-detected than those with a low score or those with minimum PCP visits, 42, 37, and 31 %, respectively (P < 0.0001). Rural patients were more likely to be symptom-detected than urban patients: 67 and 60 %, respectively (P < 0.001).

The variation in percent screen-detected cancers by RHA for all patients and screen-eligible patients only is shown in Fig. 2. Approximately 50 % of screen-eligible women were screen-detected in four RHAs, 40 % in four RHAs and only 10 % in one RHA.

Figure 3 shows the forest plot of the adjusted odds ratios from the multivariable logistic regression model in which the outcome is screen-detected cancer. RHAs that had similar estimated odds were grouped together. Patients outside of the screening-eligible age group were less likely to be screen-detected except for age group 70–74 (OR: 0.93, 95 % CI: 0.80–1.06). The likelihood of being screen-detected increased with decreasing disease stage. Compared to stage I cancer, the estimated odds ratio of being screen-detected was 1.57, 0.37, 0.19 and 0.12 for in situ, stages II, III, and IV cancers, respectively (P < 0.001). Having a comorbidity index greater than or equal to 2 was associated with a decreased odds of being screen-detected comparing to those with a comorbidity index less than 2 (OR: 0.73, 95 % CI: 0.65–0.88).

Figure 4 shows significant interaction between RHA and time period. The probability of screen-detected cancer increased in all regions from 2004–2006 to 2007–2010, however, it increased to a greater extent in RHAs 2 and 8 than other RHAs (P < 0.001).

Table 2 shows the unadjusted time to diagnosis by detection method. Three hundred forty six symptom-detected patients (4.5 % of symptom-detected patients) were excluded from the analysis because they did not have a visit with a PCP within 6 months prior to their first diagnostic test. The median (90^th percentile) time to diagnosis was 19 (70) and 21 (92) days for screen- and symptom-detected breast cancers, respectively. Cancer stage was inversely related to the diagnostic interval in symptom-detected women: as cancer stage increased from in situ to stage IV, the median diagnostic interval decreased from 49 to 13 days. The relationship was much attenuated in the screen-detected group, however; in situ cancers had a median time of 30 days while all invasive cancers (stage I through IV) had a median of about 16 days. There was also significant variation in time to diagnosis by health region: the median diagnostic interval ranged from 7 to 42 days in the screen-detected group and from 13 to 33 days in the symptom-detected group.

Table 3 shows the results of the quantile regression models for the median and the 90 % quantile for the two detection methods, based on a simplified model that did not include any interaction term.

RHA and cancer stage had the strongest associations with the time to diagnosis for both screen- and symptom-detected cancers, though the associations differed by detection mode. The estimated median time to diagnosis ranged from 4 to 37 days depending on the RHA in the screen-detected cancers (P < 0.005) and from 17 to 33 days in the symptom-detected cancers (P < 0.005), holding other factors at the reference levels. The dissimilar patterns of the adjusted median and 90^th percentile diagnostic intervals depicted in Fig. 5 illustrate the extent of regional variation. A longer median (90^th percentile) in a given RHA for a given detection mode does not necessarily translate into a longer median (90^th percentile) for the other detection mode in the same RHA.

Similarly, cancer stage showed a different pattern of variation in the screen- and symptom-detected groups. In symptom-detected cancers, the estimated median time to diagnosis decreased monotonically with increasing stage (45, 20, 14, 13 and 8 days for in situ, stage I, II, II, and IV, respectively, P < 0.005); the estimated median time in screen-detected cancers also decreased with increasing stage but had a much smaller range (15, 4, 4, 4 and 1 days for in situ, stage I, II, II, and IV, respectively, P < 0.005), holding other factors at the reference levels.

For symptom-detected patients, the effect of RHA on the length of the diagnostic interval was modified by the time period.¹ Figure 6 illustrates the effect modification of time period on RHA observed in the median time to diagnosis. From 2004–2006 to 2007–2010, the median diagnostic interval increased dramatically by 18 days in RHA 2 (P < 0.001), increased by 3 days in RHAs 1, 3 and 8 (P < 0.001), decreased by about 6 days in RHA 4 (P < 0.001), and remained largely the same in RHAs 5, 6, 7, and 9. Overall, the median diagnostic intervals were less than a month for all RHAs except RHA 2 in 2007–2010, where the median diagnostic interval was 41 days.